Expectation Formation with Correlated Variables

Abstract We experimentally study how people form expectations about correlated variables. Subjects forecast a time-series variable A. In treatment Baseline, subjects only observe past values of A. In treatment Correlated, they additionally observe a correlated variable B; A is equally predictable and has the same univariate properties in both treatments. Subjects are significantly less accurate and underreact more in Correlated, inconsistent with Bayesian learning. A structural-model estimation indicates that subjects (i) underestimate the level of correlation; and (ii) are insensitive to actual correlation. Our study provides first direct evidence of correlation neglect in the domain of expectation formation.